Agents and revisions
Manage instructions, tools, and the revision each agent runs.
Run smarter
agents. Use fewer tokens.
AgentRun is an open-source platform that supercharges your agents. It adds self-improvement, durable execution and a control plane to your favorite agent harness.
Works with Claude Code, Codex, Pi, and OpenCode.
01 SELF-LEARNING
Make your agents smarter
and cheaper with every run.
Repeated steps become reusable skills and tested workflows. Later runs reuse what works instead of planning it all again.
Learn from each run.
AgentRun turns repeated steps into proposed skills and workflows.
Test before reusing.
Proposals are checked and graded before they are accepted.
Reuse what worked.
Later runs spend fewer tokens on repeated planning and tool discovery.
LEARNING BETWEEN RUNSLearn the steps. Reuse the workflow.
First runReusable workflow
Illustrative · costs vary by run02 PLATFORM
Your choice of harness.
One platform around it.
Your harness handles reasoning and tool use. AgentRun manages its runs, workspace, recovery, and learning through one API.
Create and run an agent
import { AgentRun } from "@parcha/agentrun-client";
// Reads AGENTRUN_BASE_URL and AGENTRUN_API_KEY.
const client = new AgentRun();
const agent = await client.agents.create({
definition: {
version: 1,
name: "Research assistant",
instructions: "Research requests. Cite sources. Ask which aspect to focus on before you start.",
harness: "pi",
learning: true,
tools: ["web_search", "web_fetch", "ask_person"],
skills: [],
},
});
const started = await client.runs.start(agent.agent_id, {
input: { content: "Research acme.com" },
max_cost_usd: 1,
});
let run = await client.runs.wait(started.run_id, { timeoutMs: 600_000 });
if (run.pending?.kind === "question") {
console.log(run.pending.question);
await client.runs.message(run.run_id, { content: "Pricing and plans." });
run = await client.runs.wait(run.run_id, { timeoutMs: 600_000 });
}
console.log(run.status, run.result);Install with npm i @parcha/agentrun-client.Install with pip install parcha-agentrun. Set AGENTRUN_BASE_URL and AGENTRUN_API_KEY for your deployment.
The same client manages every run.
03 COMPUTE
Give your agents
the right compute.
Keep the agent loop separate from the computer doing the work. Connect your provider once; AgentRun handles commands, files, and cleanup through one API.
Lightweight sandboxes
Run quick scripts, inspect files, and generate small reports.
Development environments
Install dependencies, build, and test code in an isolated Linux environment.
High-resource sandboxes
Configure more CPU and memory for larger datasets and batch processing.
04 DURABLE EXECUTION
Keep your progress.
Even when a run stops.
AgentRun saves the conversation, files, and completed tool results. After an interruption, recover the run from saved state instead of rebuilding its context.
One recovery layer
Keep the same recovery API across harnesses. No need to rebuild persistence around each one.
Saved work
Keep completed tool results alongside the conversation and workspace.
Restart recovery
A new process reopens the saved state and continues the run.
Continue a stopped runTypeScript client
import { AgentRun } from "@parcha/agentrun-client";
const client = new AgentRun();
// A stopped run keeps its conversation and files.
await client.runs.message("your-run-id", {
content: "Continue with the saved context.",
});
const run = await client.runs.wait("your-run-id");
console.log(run.status, run.result);New turn. Same conversation and files.
PLATFORM FEATURES
Give your agents everything they need to get work done.
Connect tools. Recover interrupted runs. Turn repeated work into reusable skills.
Turn repeated work into skills
Turn repeated steps into reusable skills and tested workflows that later runs can use.
Read the docs
Choose your agent harness
Choose Pi, Claude Code, Codex, or OpenCode. Keep the same API for agents and runs.
Read the docs
Pick up where runs left off
Recover interrupted runs with saved conversation, files, and recorded tool results.
Read the docs
Manage every agent in one place
Manage multiple agents, version their definitions, and inspect every run from one console.
Read the docs
Build workflows you can reuse
Combine tools, decisions, and agent steps into workflows you can inspect, test, and reuse.
Read the docs
Connect the tools your agents need
Connect built-in tools and remote MCP servers, with explicit tool access for each agent.
Read the docs
Keep the files your agents create
Give each run its own files. Upload inputs, inspect outputs, and continue with saved work.
Read the docs
Run it on your infrastructure
Run the Apache-2.0 platform on your own machine or server with your own model keys.
Read the docs05 AGENT MANAGEMENT
Manage all your agents
in one place.
See what your agents are doing, find the runs that need attention, and inspect their results through one console and API.
Runs and results
Monitor progress, respond to requests for input, and inspect files.
Learning notes
Review what an agent learned and the context available to its next run.
PLATFORM COMPARISON
Why AgentRun?
Choose your harness. Own your infrastructure. Turn repeated work into reusable skills and tested workflows.
Included
Not offered
Partial / different scope
Not documented
Scroll to compare platforms
| What you get | Open sourceAgentRun | AnthropicClaude Managed Agents | OpenAIAgents API | Google CloudGemini Enterprise Agent Platform |
|---|---|---|---|---|
| Open source & self-hostableFull platform, including the console | Apache 2.0 | Managed service | Managed service | Managed service |
| One API across harnessesChoose how your agents run | Pi · Claude Code Codex · OpenCode |
Claude only | Codex only | Custom integration |
| Learns skills & workflowsFrom completed runs | Built-in, evaluated learning | Memory learning Preview |
Not documented | Skill optimization Experimental |
| Durable executionRecover interrupted work | Built-in recovery | Durable sessions | Durable sessions | Opt-in resume |
| Run multiple agentsManage concurrent work | Shared API & console | Agent teams | Parallel subagents | Multi-agent workflows |
06 DEPLOYMENT
Open source.
Or hosted for you.
Learning, recovery, and the console are all open source. Run AgentRun with your own API keys, or choose a hosted deployment.
Built by the team behind Grep AI, whose platform powers business process agents at Worldpay, IG.com, Citizens, and Airwallex. Self-host AgentRun with your own keys, or let us host it on AgentRun Cloud.
FAQ
Frequently asked questions
How learning works, where the savings come from, and what stays under your control.
What does AgentRun add to my agent harness?
AgentRun is the platform around your chosen harness. Define agents, run them through a common API, and manage their revisions, run history, recovery, and learning in one place. The harness still handles the agent’s reasoning and tool use.
Is AgentRun open source? How does it relate to Grep?
AgentRun is open source under Apache-2.0, built by the team behind Grep. Learning, recovery, and the console are included. You can self-host with your own model API keys; model and infrastructure costs are yours. Grep is our managed platform for business agents. AgentRun Cloud is our hosted AgentRun offering, and is optional when you self-host.
How does self-learning work?
When learning is enabled, the agent reflects on its runs and proposes notes, reusable skills, and workflow drafts. Accepted changes are versioned and available to later runs, so repeated steps can be reused instead of planned again. This updates the agent’s saved knowledge and procedures, not the model’s weights.
How does that reduce token usage?
An agent can spend tokens rediscovering an API, planning the same request, or deciding how to validate a result. Reusing a skill or workflow can skip that repeated reasoning. The agent still uses models and tools where needed. Potential savings come from reducing repeated work.
How much will I save?
Savings depend on how much work repeats. The launch post reported a 90% cost reduction for one compliance workload. Results vary, and reflection and grading also add cost. Measure total spend and answer quality on your own workload, including the cost of learning.
What happens when a decision is uncertain?
You define the threshold and the fallback in your workflow. Uncertain results can go to an agent for further investigation or trigger escalation for review. A confidence score is a routing signal, not a guarantee of accuracy. Choose thresholds using evaluated examples and the consequences of getting the decision wrong.
Do I need Jev to use AgentRun?
No. Jev is an optional integration for focused decisions in workflows you author, such as choosing a category or returning a score. Agents and self-learning work without it. Automatically learned workflows currently use supported model and tool steps; they do not automatically turn into Jev decisions.
Do I have to turn my whole agent into a workflow?
No. Keep open-ended tasks agentic and make repeatable steps reusable. Workflows can include agent investigation at defined points, so you can automate routine paths while deciding how exceptions should be handled. You control where a workflow continues, retries, or asks for help.
How are learned changes checked?
Proposals pass structural checks and an AI grader before they can be reused. Changes that widen trust, loosen checks, or direct money or data require owner approval. You can inspect changes, edit notes, or retire learned material. These checks help filter proposals; you still need to evaluate results on your tasks.
Which harnesses can I use?
Pi, Claude Code through the Claude Agent SDK, Codex, and OpenCode. Choose a harness in your AgentRun agent definition while keeping shared run management and learning. Each harness needs its own configuration, and available models and tools can differ.
Can I run AgentRun on my own infrastructure?
Yes. Self-host the open-source platform locally or on your own server and bring your model API keys. Agent definitions, runs, and learned material live in your configured data directory. Model calls and hosted browser sessions still go to the providers you configure.